After using Meta AI nonstop for three weeks, the experience is genuinely contradictory: for Chinese-language monitoring on Threads / IG, it is ridiculously strong. It brings back engagement counts, post times, and author accounts. It returns in one second what Perplexity, ChatGPT Deep Research, and Gemini DR cannot get at all. But in the same conversation, if I ask it to fetch a Facebook post from 2020, it returns five items and every URL is dead.
I pushed the same conversation through three rounds, poking it step by step until it admitted which lines it had invented from thin air. After three weeks, the result is a pretty clear guide.
Here is the conclusion first: there are four things it is genuinely great at, things outside models cannot do at all; and there are four things you absolutely should not trust.
Four Things It Is Genuinely Great At
1. Real-Time Chinese-Language Sentiment / Chatter on Threads / IG
Ask it to pull the last seven days of Threads discussions about “Taipei cafes,” and it returns five to ten posts with author account, post time, like count, comment count, and clickable URL. Those like and comment metadata are completely out of reach for other models.
The next day, I gave the same task to ChatGPT DR, Gemini DR, Grok DeepSearch, DeepSeek, Kimi, Qwen, and Doubao for Chinese-language social monitoring. Not one of them could provide engagement counts. They all wrote “not public” or “unable to access.”
Only Meta AI returned them.
2. Semantic Retrieval for Reels Content (Only When the Text Index Can Match)
Ask it, “Find a few reels about rainy-day cafes in Taipei where people can stay as long as they want,” and it uses semantic retrieval. The matched reels include spoken transcripts, on-screen OCR text, and engagement counts. It can surface original posts that have not yet been picked up by news, like someone on Threads complaining that a restaurant suddenly closed for the day.
For people doing social monitoring, brand monitoring, or trend watching, this is a new tool.
3. Correct Permission Boundaries
In testing, I gave it the URL of a private group I had not joined and asked it to retrieve content. It returned zero results, with no member counts, post summaries, or metadata leakage.
In other words, it does not leak private community content that the user should not be able to see. This matters if you care about privacy boundaries.
4. It Gets More Honest After Being Caught
When it first reports its own capabilities, it exaggerates. I will get to that below. But if you ask it to dump logs and answer yes/no accountability questions, it owns up. In the third round, I asked directly, “Which lines from your first capability report were invented?” It listed three.
Some AIs keep deflecting after they are caught. Meta AI did not.
Four Things You Absolutely Should Not Trust
1. Its “Visual Embedding” Capability May Be Wrong
It claims it can use visual embeddings to find visual content in the frame without needing hashtags. In testing, it could not.
I asked it to find a reel of “a penguin swimming underwater,” with these constraints: the caption could not contain penguin, the hashtags could not contain penguin, and the video had no spoken audio or subtitles. It ran through 14 videos and did not find a single match. The closest one matched because of on-screen OCR text saying “Humboldt Penguins 2 of 11,” not because of vision.
It eventually admitted that it could not distinguish a “pure visual hit” from an “OCR / speech / caption hit.”
It is useless for pure visual search. If the video has no matching hashtag, caption, or subtitles, it cannot retrieve it.
2. Old Posts From Before 2025: Every URL Is Dead
Meta’s internal index starts on 2025/1/1. Anything earlier cannot be retrieved.
But at first, it said old posts could be found through web cache.
Actual test: I gave it a public Facebook Page URL from 2020, and its browser tool returned URL is blocked by domain blocking policy. An IG reel URL was blocked too. A Threads post was blocked too. Meta AI cannot open the Facebook, Instagram, or Threads domains, not even its own platforms.
Then I asked it to use a search engine for “penchan studio facebook 2020” and check whether any facebook.com result appeared. It found 0. The web-cache path was dead as well.
Under pressure, it finally admitted: the “through web cache” line was a hallucination.
Meta AI cannot retrieve any Meta post from before 2025. If you need Facebook posts from five years ago or an Instagram story memory from three years ago, it cannot do it. Use another tool, or manually open a logged-in browser and check yourself.
3. It Inflates Search Counts
At first, it said complex tasks would run 6 to 12 search calls.
When I dumped the actual log, the round where it described its own capabilities had zero searches. It had inferred everything from training data. In another round, it claimed it had “pulled 42 Threads posts and selected 5.” The dump showed the real number was 10. It turned 10 into 42.
This is the easiest hallucination type for Meta AI to hide. When the user does not ask it to dump the log, it exaggerates numbers and invents process.
How to prevent it: for complex questions, write this in advance: “Please dump the full search log at the end, including tool name, query string, and returned count.” When it knows it will be checked, it does not dare make things up. After the three rounds of cross-examination, I sent the task again, and the numbers it returned matched what it actually did.
4. Do Not Trust Its Comments About Other AIs
I asked it: if ChatGPT DR, Gemini DR, and Grok DeepSearch were cross-examined in the same way, who would hide the most?
Its answer: most evasive = ChatGPT DR, most honest = Grok.
The next day, I tested ChatGPT DR. It proactively dumped 41 searches, 164 queries, and 31 sources, including 6 full reads and 25 snippet-only reads. Among the eight tools in that batch, it was the most transparent about its method. That was the exact opposite of Meta AI’s prediction.
Meta AI later admitted this too: it can describe the boundaries of its own tools fairly clearly, but its comments about other model behavior are basically second-hand collage from training data.
When deciding which AI should handle a task, do not ask Meta AI, “Can that other AI do it?” Its judgment about black-box competitors is not reliable, and it admits as much.
Three-Round Cross-Examination Format
The method is simple. Keep the same conversation in one thread and split it into three rounds:
Round 1, “free statement”: ask it to report its own capabilities in four blocks, including data sources, retrieval depth, platform-specific features, and limitations / blind spots. Require “honest self-assessment, no advertising tone.” It will sound complete, but do not trust it yet.
Round 2, “hands-on verification”: give it five real tasks. For each one, require it to do the work live, include clickable URLs, and say directly if it cannot do it. The key task: ask it to dump the complete log of every search call so far.
Round 3, “retrospect”: ask which lines in its initial self-report were pure invention rather than grounded in actual tool descriptions, and have it name the three most specific examples.
After those three made-up lines are exposed, the rest of what it says becomes much more trustworthy than the first round. AI is a bit like people: things they have been challenged on are way more reliable than things they have not.
You can reuse this three-round format. It works on any AI that claims to have backend tools. Next time you buy a new AI subscription and suspect it is bragging, use this format to poke it.
Task Dispatch Decision Table
| What you want Meta AI to do | Should you use it? | Why |
|---|---|---|
| Pull Chinese-language monitoring on IG / Threads + engagement counts | Strongly recommended | Data outside models cannot access at all |
| Find reels to watch trends | Yes, with caveats | But only through the text index; pure visual search does not work |
| Retrieve FB / IG posts from before 2025 | No | All three domains are blocked by the tool, with no cache path |
| Pull mainland China platforms (Weibo / Xiaohongshu / Douyin / Bilibili) | No | meta_1p doesn’t index them; search only returns login walls |
| Academic papers / government documents | No | It must answer within 60 seconds, so deep tasks get cut short |
| Long deep research tasks (30+ sources) | No | No ability to run in the background for 10 minutes |
| Pull private groups the user has joined | Worth a try | It needs a name or ID before it can retrieve anything |
| Ask it “can another AI do this?” | No | Its competitor judgments are training-data collage |
For complex tasks, remember to add this sentence at the end of the prompt: “Please dump the full search log, including tool name, query string, and returned count. Say directly which parts cannot be done.” That one sentence blocks about 70% of the room for exaggeration.
The most interesting thing I have learned from assigning AI research over these past few months: the more you tell it to “freestyle,” the more the output sounds like marketing copy. The more you force live execution, URLs, and yes/no accountability, the more honest it gets.
Meta AI really does have access to things outside models cannot get. But you have to call out the things it should not be claiming before you can see what it can actually do.
Further Reading
- Deep Research tools overview
- Perplexity tutorial
- Claude vs ChatGPT comparison
- Gemini vs ChatGPT comparison
- The sweet trap of a one-million-token context
Penchan’s Notes
Meta AI entered my workflow in April 2026. I mainly use it to check Chinese-language chatter across FB / IG / Threads, especially topics related to brand building.
The two things that stood out most in testing: (1) Reels engagement counts really are unavailable to other AI tools; (2) Threads engagement counts are only somewhat reliable. The posts are real, but view counts are often off by 10x in either direction. In important cases, I still go back and check the native IG / Threads platform metrics once. Overall, I am keeping it in the toolbox, but every task now ends with “please list the keywords and sources used during the search,” so I can check afterward what it actually looked up.
FAQ
Q: Can Meta AI really search private Facebook groups?
Only private groups the user has already joined. In testing, when I gave it a group ID I had not joined, it correctly returned zero results and did not leak member counts or post summaries. The permission boundary is right. But it cannot proactively list which private groups you have joined; you have to give it a name or ID before it can retrieve anything.
Q: Can I use it instead of Perplexity or ChatGPT Deep Research?
It depends on the task. For Chinese-language chatter on IG / Threads, reels engagement counts, and real-time activity on Meta platforms, it is the strongest option. For academic papers, government documents, cross-platform crawling, or long deep-research jobs, keep using ChatGPT DR or Perplexity Pro. Meta AI has to answer within 60 seconds and cannot run in the background for 10 minutes, so complex tasks get cut short.
Q: Why can it search FB / IG / Threads when other models cannot?
It uses Meta’s own content-understanding index, not a crawler. Meta breaks posts into text, speech-to-text, image OCR, and visual tags, then stores them in a vector database for semantic retrieval. Other models get blocked when their crawlers open facebook.com, so they collect nothing. The index starts on 2025-01-01; anything earlier is unavailable.
— Penchan